GenAI and copyright: three key cases

The rise of Generative Artificial Intelligence (GenAI) has sparked a global debate on how copyright laws impact its use. A core question is whether the use of copyright protected works (Inputs) to create large language models (LLMs) and the subsequent generation of content (Outputs) by users of the LLMs constitute copyright infringements. Allied to this is whether such uses can be shielded by defences or exemptions like fair use or permissible text and data mining or by deeming Outputs as “transformative”. Amidst this uncertainty, three significant cases are proceeding through the courts in the USA, UK and the EU that may offer some answers to these questions (respectively, Bartz v Anthropic (US), Getty Images v. Stability AI (UK), and Like Company v Google Ireland (EU)). This article reviews the progress of each case and when we may expect some final definitive answers from each jurisdiction.  

USA: Bartz v Anthropic PBC

This marks the first significant American court ruling on how the US doctrine of fair use applies to GenAI. Three authors have sued Anthropic in the Californian courts for allegedly copying their copyrighted books to train its GenAI model, called Claude. The claim focuses on Inputs. There is no claim that the Outputs generated by Claude infringe the authors’ works.  

In June 2025 Judge Alsup granted the defendant, Anthropic, partial summary judgement over the authors’ claims (in other words these aspects of the claim were so clear that they did not need to go to trial). The Judge decided that where Anthropic had purchased and scanned books for the subsequent training of its GenAI model then this constituted fair use under the US Copyright Act. The Court found that Anthropic had simply replaced the print copies with scanned versions and this had not reduced demand for the original books. The Court held that this use was “spectacularly” transformative, akin to human learning, and did not harm the market for the original works. The Court drew analogies to previous fair use rulings such as Authors Guild Inc v HathiTrust. In that case, it was decided that the creation of a full-text searchable database through the large-scale digitization of books, coupled with the provision of accessible formats for print-disabled individuals, was a transformative application of the works within the meaning of the US doctrine of fair use.

However, Judge Alsup refused Anthropic summary judgement on that part of the case where the authors relied on the downloading of pirated copies of books. The Judge said the use of pirated copies had reduced demand for the original books and he therefore doubted it could be fair use. This question will now proceed to a full trial likely to take place in December 2025.

Therefore, this ruling highlights how US courts might draw a brightline between LLMs based on lawfully acquired and pirated training data, and that the fair use exception may protect transformative AI training under some circumstances.

The finding by the Judge that what happens in training and creating LLMs is akin to human learning and adaptation is particularly controversial. Some argue that a predictive model like an LLM cannot be equated with human learning and reasoning.   

UK: Getty Images v Stability AI

UK-based Getty Images has brought multiple claims for copyright, trade mark and database rights infringements against Stability AI, the developer of the text-to-image and image-to-image GenAI model called Stable Diffusion. Getty alleges that Stability copied millions of images on Getty’s site to train Stable Diffusion without their permission (the Input Claim). Getty further claims that the outputs of the AI were similar to copyrighted images belonging to Getty (the Output Claim), with some even having Getty’s watermark on it which in turn constituted infringement of Getty’s trade mark rights and passing off (the Brand Claims). Finally, Getty alleges that by making Stable Diffusion available to users in the UK, Stability is importing an infringing article into the UK (the Importation Claim).   

The High Court trial took place in June 2025. A decision is expected in the autumn. However, in closing arguments at trial Getty withdrew both the Input Claim and the Output Claim. Getty explained it was abandoning the Input Claim because it could not prove on the evidence presented at trial that any development of Stable Diffusion had actually taken place in the UK (rather than in the USA). With regard to the Output Claim, Getty said that Stability had adjusted its model such that no Outputs generated by users would now reproduce a substantial part of Getty’s original works in a legally meaningful way. Therefore, Getty said this part of the claim had become redundant and the request for injunction to prevent this happening was now pointless. However, the High Court may still need to consider the Output Claim and the reasons for Getty dropping it when it comes to the question of deciding who pays the legal costs associated with that issue (in other words, would Getty ever have been able to prove the Output Claim?).  

However, the Brand Claims and the Importation Claim are still live. The Importation Claim is particularly important. If the High Court rules that making Stable Diffusion available in the UK constitutes the importation of an infringing “article” into the UK, then that is likely to lead to an injunction requiring Stability not to make Stable Diffusion available to anyone in the UK. This would be very significant for the GenAI industry because it would establish that the UK can be ringfenced from using any infringing GenAI models. The ruling will turn on whether an intangible item like Stable Diffusion can be an “article” for these purposes (earlier cases have concluded that physical items such as footwear or CDs are “articles”, but there have been no cases to date covering something intangible like software). Stability tried to strike out the Importation Claim earlier in the proceedings, but the Judge ruled that it was sufficiently arguable to go to trial. So, the point looks to be on a knife-edge and is in our view the most significant issue to watch out for when the ruling is handed down, likely in the Autumn.        

EU: Like Company v Google Ireland

Finally, the Like Company case pending before the CJEU seems set to be the first time the EU’s highest Court will rule on whether GenAI models infringe copyright law. In April 2025, the Budapest District Court referred four preliminary questions to the Court of Justice of the European Union (the CJEU) after Like Company, a Hungarian news publisher, alleged that Google’s AI model, Gemini, reproduced and summarised its copyrighted news articles without their permission.

The CJEU has been asked to determine whether training an LLM on protected copyright works constitutes reproduction under Article 2 of the Copyright and Information Society Directive 2001 (the InfoSoc Directive) and whether AI-generated outputs constitute a communication to the public under both the InfoSoc Directive and the Digital Single Market Directive 2019 (DSM Directive). Further, the CJEU has been asked to consider whether the text and data mining exception under Article 4 of the DSM Directive applies to commercial GenAI training. The case raises novel issues around the scope of exceptions and the attribution of liability for GenAI outputs that mimic copyrighted works. A decision is expected no earlier than 2027, but its impact will likely define how GenAI developers can use copyrighted content in the EU.

Conclusion

Each of these three cases is significant in its own right, but each illustrates the differing legal regimes that apply to GenAI and copyright globally. At the heart of American cases are questions of fair use, a doctrine which is much wider in the USA in affording a defence than in other jurisdictions. However, is it wide enough for example to accommodate all stages of the development and use of GenAI – acquisition of content, training and use? The Getty case pending before the English courts demonstrates both the importance of establishing where allegedly infringing acts of acquiring content and training AI may have occurred, but also the potential long-arm effect of national copyright laws that may prevent the use of GenAI by users when made available in a particular country. In other words, even if the GenAI is lawfully created in one country (eg in the USA due to its wide concepts of fair use) does not mean that it can be offered lawfully in other countries. Finally, the CJEU case looks set to test important fundamental concepts around EU laws impacting Inputs, Outputs and the Text and Data Mining exception. The potential for differing outcomes and orders in these cases, and hence a fragmented global approach, is obvious.

Many thanks to Laura Biro, current paralegal in the team, for her help preparing this article.

This publication is a general summary of the law. It should not replace legal advice tailored to your specific circumstances.

© Farrer & Co LLP, July 2025